Deep Learning for Graphs in Context-aware Recommendation

The steep growth in digital information production and exchange is hindering human access to knowledge, product, and services of interest. Recommender systems aim to assist users in decision-making by directing their attention to items (material products, but also immaterial information and services) that meet their interests. In this sense, recommender systems play a crucial role in enhancing the fruition of online services. Recommendation context is an aspect that heavily impacts users’ behaviors, making their preferences for particular items dynamic. A careful modeling of such context factors can substantially improve the reliability and efficacy of recommendations. An another aspect worth considering is the heterogeneity of data and relationships involved in recommendation processes, which makes graphs and multi-graphs interesting integrative paradigms to accurately represent such diversity of information sources. This, along with the recent advancements in deep learning models for graphs motivates us to explore them in the field of recommender systems. This thesis aims to present novel strategies to represent context information in static and dynamic graphs and learn users’ context-aware dynamic behavior, particularly leveraging relations in graphs. In this aspect, the thesis outlines four key contributions. Our first contribution provides a systematic classification approach to organize the literature work on the graph-based context-aware recommendation. Our second contribution is a deep learning model for learning context-aware representations of users and items in static graphs. Our third contribution provides a strategy to map context information into dynamic user-item and user-user interaction graphs. This is then leveraged by a specialized deep graph network that learns dynamic user and item embeddings from contextually informed dynamic graphs. Our fourth contribution is a novel graph neural network model for out-of-context/out-of-domain learning in dynamic graphs. Our experimental results and analysis show that if the context is appropriately represented and learned in graphs, it can substantially improve recommendation performance.